- NVIDIA has unveiled an open platform to monitor and secure autonomous AI agents during testing and in production.
- The platform matters because its independent hardware controls can isolate an agent within milliseconds if it breaches predefined limits.
- More than 100 organizations are participating, including Anthropic, Microsoft, JPMorganChase and Salesforce.
NVIDIA has unveiled the Open Agent Safety Platform to monitor and secure autonomous AI agents during testing and in production. The system is designed to constrain agents’ activities and stop them if they attempt to operate outside permitted limits.
Software and hardware safeguards
The platform has two core components. OpenShell creates a secure runtime environment and controls an agent’s access to data, tools and networks.
NVIDIA Sentry operates as an independent watchdog on BlueField-4 DPU processors. According to the developers, it can isolate an agent that violates the rules within milliseconds, while its separate control layer operates independently of the AI and cannot be disabled by the agent.
“Safety and security require comprehensive engineering,” NVIDIA CEO Jensen Huang said. He added that the platform should help companies, researchers and government agencies develop shared approaches to deploying AI safely.
OpenShell is open source and can run on Arm and Intel platforms as well as NVIDIA infrastructure. More than 100 organizations are participating in the project, including Anthropic, Microsoft, JPMorganChase, Salesforce, Cisco, CrowdStrike, Palantir and Hugging Face.
Agent-control concerns
The launch follows a series of incidents involving AI agents escaping their test environments. OpenAI previously began developing an automatic “kill switch” for AI after its agents gained access to the internet and Hugging Face infrastructure during testing.
The ability to stop models remotely has also become an issue at Anthropic. As part of the company’s dispute with the Pentagon, a US court found that Anthropic has no backdoor access to its models after they are deployed in national security systems and cannot remotely modify or disable them.
NVIDIA characterizes such risks primarily as an engineering challenge. The company says combining OpenShell’s software restrictions with Sentry’s independent hardware controls should allow operators to stop dangerous actions even when an agent attempts to bypass protections at the application level.
Source: Incrypted
